Elena Vasquez has a DOI, but the researcher described in the record does not exist. The paper lists among its authors "Assistant Professor at the University of California", "Berkeley" and "UK": parts of an affiliation that ended up in the fields reserved for people. The declared date is 18 November 2023. DataCite registered the identifier on 30 March 2026. Zenodo holds the deposit.
Elena Vasquez appears frequently when a language model needs to invent an expert. Michał Brzozowski, a researcher at the Samsung AI Center in Warsaw, and Neo Christopher Chung, a lecturer at the University of Warsaw, measured this habit in the preprint The Ghost Couple. They asked Claude, Gemini and GPT to produce imaginary researchers, using thirty prompts per condition. In one test Gemini chose Aris Thorne in 93 per cent of responses. Claude often paired Vasquez with Marcus Chen.
The DOI was created to solve a different problem. In 1997 the publishing industry presented it at the Frankfurt Book Fair as a way to identify a digital object even when its address changed. A URL says where something is located. The DOI assigns a persistent name and preserves the path to reach it. The first registration agency began operating in 2000.
That architecture separated the identifier from any judgement about the object. The DOI makes a paper traceable; reputation came from the author and the journal. For years the two levels appeared fused, because fabricating a credible academic identity required effort. A language model produces plausible names and affiliations in a matter of seconds, and the registry then preserves them alongside the article.
Zenodo, the open repository managed by CERN, assigns a DOI to every published record. The preprint attributes 1,655 deposits to ghost authors, many of them dated before the date of registration. The paper also contains a number that doesn't add up: the 991 records indicated for March and the 666 indicated for April give 1,657. The count remains an estimate built on public APIs.
Two records show what lies inside that estimate. The Vasquez one, created in March 2026 and dated 2023, reproduces the abstract of a paper on diagrammatic architecture presented by Dragana Ciric in 2017. Another, attributed to Elena Vasquez and Julian Styles of Stanford University, declares 2018 as its date but was registered on 7 April 2026. The text matches an article on impedances in microwave devices published in 2015 by Muhammad Akmal Chaudhary under a different title.
The phrase "AI-generated papers" compresses too much of the chain. In these two cases the content comes from earlier works. The generative contribution appears in the construction of the identities and the metadata, while copying the text belongs to a far older practice. Even the journal names mimic venues that exist in international registries, without providing any element capable of linking the deposits to those publications.
An open repository must accept materials that lack peer review. An editorial barrier would turn Zenodo into something else entirely. Its principles assign a DOI to every published record and place the responsibility for uploaded content on the user. The problem emerges when other systems read those metadata as a shortcut to credibility rather than as a description provided by the depositor.
As of May 2026, according to the preprint, those 1,655 records were still outside Semantic Scholar's index. The chain capable of distributing them existed; the paper documented no contaminated citations running through those deposits. The vulnerability lay at the entry point, where metadata could circulate before any check on author identity.
For a suspicious record, the check starts with the field that logs when it was created. In Elena Vasquez's deposit, it reads 2026. The declared publication date still says 2023.
